Black Forest Labs Releases FLUX.2 [klein]: Integrated Flow Models for Interactive Visual Intelligence

Black Forest Labs releases FLUX.2 [klein]a family of integrated graphics models that target interactive virtual intelligence in consumer hardware. FLUX.2 [klein] extends the FLUX.2 line with a second sub-generation and editing, integrated text-to-image and image-to-image architecture, and deployment options ranging from local GPUs to cloud APIs, while maintaining state-of-the-art image quality.

From FLUX.2 [dev] in interactive virtual intelligence

FLUX.2 [dev] is a 32-billion-parameter dynamic flow for generating and editing text-based images, including multi-reference image architectures, and works primarily on data center-class accelerators. Optimized for high quality and flexibility, with long sampling schedules and high VRAM requirements.

FLUX.2 [klein] takes the same design direction and condenses it into small tuned flux converters with 4 billion and 9 billion parameters. These models are scaled down to very short sampling schedules, support parallel text-to-image and multi-reference editing operations, and are optimized for response times of less than 1 second on today’s GPUs.

Family model and skills

FLUX.2 [klein] the family consists of 4 main types of open mass with a single structure.

• FLUX.2 [klein] 4B
• FLUX.2 [klein] 9B
• FLUX.2 [klein] 4B Foundation
• FLUX.2 [klein] 9B base

FLUX.2 [klein] 4B and 9B are distilled and direction models. They use a 4-step indexing process and are positioned as the fastest options for productivity and interactive workloads. FLUX.2 [klein] 9B combines the 9B flow model with the 8B Qwen3 text encoder and is described as the best submodel on the Pareto frontier for quality versus text-to-image delay, single-reference editing, and multi-reference processing.

Base variants are undiluted versions with longer sampling schedules. The literature lists them as basic models that preserve the perfect training signal and provide high output diversity. They are designed for fine-tuning, LoRA training, research pipelines, and post-customization workflows where control is more important than minimal latency.

All FLUX.2 [klein] models support three main functions in the same structure. They can generate images with text, they can edit a single input image, and they can do multiple references and editing where several input images and information jointly define a target output.

Latency, VRAM, and limited variance

FLUX.2 [klein] The model page gives estimated end-to-end times for the GB200 and RTX 5090. FLUX.2 [klein] 4B is the fastest version and is listed at 0.3 to 1.2 seconds per image, depending on the hardware. FLUX.2 [klein] 9B targets about 0.5 to 2 seconds in high quality. Base models require fewer seconds because they work with 50-step sampling schedules, but offer more flexibility for custom pipelines.

FLUX.2 [klein] The 4B model card says the 4B is equivalent to 13 GB of VRAM and is suitable for GPUs such as RTX 3090 and RTX 4070. FLUX.2 [klein] The 9B card reports a requirement of around 29 GB of VRAM and target hardware such as the RTX 4090. This means that a single high-end consumer card can handle the distilled variant at full sample resolution.

To extend access to more devices, Black Forest Labs is also releasing FP8 and NVFP4 versions of all FLUX.2 [klein] different, developed together with NVIDIA. The FP8 benchmark is described as up to 1.6 times faster up to 40 percent lower VRAM usage, and NVFP4 up to 2.7 times faster up to 55 percent lower VRAM usage on RTX GPUs, while maintaining the same core power.

Benchmarks are compared to other graphics models

Black Forest Labs tests FLUX.2 [klein] using Elo’s text and image comparison styles, single reference editing, and multiple reference functions. Performance charts showing FLUX.2 [klein] on the Pareto frontier of Elo score versus latency and Elo score versus VRAM. Definition is FLUX.2 [klein] it matches or exceeds the quality of Qwen-based image models in terms of latency and VRAM, and that it outperforms Z Image while supporting integrated text for image processing and multi-reference editing in a single build.

The basic variant trades some speed for full customization and fine-tuning, which fits its role as basic testbeds for new research and some domain-specific pipelines.

Key Takeaways

• FLUX.2 [klein] is a family of fixed-flow converters with variants 4B and 9B that support text-to-image, single-image editing, and multiple references to a single integrated structure.
• FLUX.2 [klein] The 4B and 9B models use 4 sampling steps and are optimized for sub-second processing on a single modern GPU, while the uncapped Base models use longer schedules and are intended for fine tuning and research.
• Quantized FP8 and NVFP4 variants, built with NVIDIA, provide up to 1.6 times speedup with approximately 40 percent VRAM reduction for FP8 and up to 2.7 times speedup with approximately 55 percent VRAM reduction for NVFP4 on RTX GPUs.

Check it out Technical details, Repo again Model weights. Also, feel free to follow us Twitter and don’t forget to join our 100k+ ML SubReddit and Subscribe to Our newspaper. Wait! are you on telegram? now you can join us on telegram too.

Michal Sutter is a data science expert with a Master of Science in Data Science from the University of Padova. With a strong foundation in statistical analysis, machine learning, and data engineering, Michal excels at turning complex data sets into actionable insights.